Geo-Distributed Driving Maneuver Anomaly Detection

Abstract

Auto-Encoder has been widely applied to anomaly detection areas. In this paper, we present a geo-distributed driving maneuver anomaly detection system based on auto-encoder. The auto-encoder is trained by using the normal driving data, so it memorizes the feature of normal driving pattern. The well trained auto-encoder is able to work as a classifier during the detection phase, it will tell whether the input data is normal or abnormal. To further improve the detection accuracy, we divide a city into a set of sub-regions by maximizing the spatial contrast within the same sub-region and minimizing the spatial contrast among different sub-regions. To examine performance of the proposed system, we evaluate it using a large dataset of GPS trajectories. The experiment results show our system achieves high detection accuracy.